DiGCT: Synergistic Physics-Data Constrained Diffusion Model for Surface Thermal Management of Press-Pack IGCTs

Official implementation for "Synergistic Physics-Data Constrained Diffusion Model for Surface Thermal Management of Press-Pack IGCTs". This project enables a diffusion-based digital twin model to monitor, evaluate, and optimize the surface temperature distribution of press-pack IGCTs.

✨ Highlights

• Synergistic Physics-Data Integration: The GCT surface temperature reference is constructed by interpolating analytical predictions and real-time temperature measurements. This synergistic integration compresses the physical mechanisms and empirical observations into a geometric representation.
• Heuristic Physics-Constrained Refinement: The proposed diffusion model iteratively refines the residual error of the reference, generating high-fidelity GCT surface temperature distribution following specific regulation and consistency requirements.
• Gradient-Based Temperature Optimization: An online optimization strategy is developed to regulate GCT surface temperature distributions, supporting diverse metrics such as maximum value, mean value, and spatial variance.
• Specialized Dataset IGCT X: The first dataset tailored for surface thermal management of press-pack IGCTs is introduced. It contains GCT surface and side temperature data in pairs, considering multiple physics coupling effects and varied system parameters.

🧩 Setup Guideline

Please meet the package requirement of assets/requirement.yaml.

conda env create -n DiGCT -f requirement.yml

In general, the following dependencies should be installed

• Python >= 3.12
• PyTorch >= 1.6.0

🔥 Quickstart

💪 Model Training

• Adjust the key parameters for model training in configs/config_model.yml

  • training: on-off switch for training
  • generate_sample: on-off switch for sample generation after training
  • physics_constraint: on-off switch for physics-constrained denoising refinement

• Train model. The training results should appear in the folder results

python model.py -config configs/config_model.yml

✍️ Model Testing

• Adjust the key parameters for model testing in configs/config_model.yml

  • testing: on-off switch for testing
  • test_path: path of result folder
  • calculate_metric: evaluate the model performance based on the generated samples
  • sample_metric: evaluate the model performance through sampling process

• Test model. The training results should appear in the corresponding testing folder

python model.py -config configs/config_model.yml

🙏 Acknowledgement

The project is built based on the following repository:

• hojonathanho/diffusion
• ximinng/LLM4SVG

We gratefully thank the authors for their wonderful works.

📋 Citation

If you use this code for your research, please cite the following work:

@ARTICLE{11567995,
  author={Yang, Xiao and Xiao, Yu and Li, Tianchen and Yang, Dongsheng},
  journal={IEEE Transactions on Industrial Informatics}, 
  title={Synergistic Physics-Data Constrained Diffusion Model for Surface Thermal Management of Press-Pack IGCTs}, 
  year={2026},
  pages={1-11},
  doi={10.1109/TII.2026.3693363}}

☎️ Contact

If you have any questions, please contact the authors at x.yang2@tue.nl

©️ License

This work is licensed under the MIT License.